Multi-Scale Bushfire Detection From Multi-Modal Streams of Remote Sensing Data

Real-time wildfire detection with semantic explanations

“…Our framework is relevant to existing systems in the sense that each component of our framework can be reused. For example, the semantic annotation pipeline can be integrated to improve the classification features of existing systems based on online classifiers (Phan et al, 2020). Moreover, our complex event processing component can encode rule-based patterns if domain knowledge for wildfire detection is known before-hand (Boubeta-Puig et al, 2021).…”

“…Model-intrinsic No Post-processing Low/Medium 1 e.g. (Phan et al, 2020;Sousa et al, 2020) 2 e.g. (Xu et al, 2010;Xu & Zhong, 2017) 3 e.g.…”

“…Most wildfire detection models inherently lack human-understandable explanations for their output. On the one hand, complex models like classifiers (Phan et al, 2020) rely on a set of raw features, which are too low-level and noisy to facilitate human understanding. Even with the advance of deep learning methods in which the latent features are learned automatically, these features still lack the ability to generate rich semantic explanations, especially with streaming data.…”

“…Raw data streams have different characteristics. For example, the sampling rate of a data stream is often fixed, but different data sources can have different sampling intervals (Phan et al, 2020) (e.g., 15 min for satellite streams and 5-60 min for weather streams). We call all of these different data streams heterogeneous data streams, and they are the input for our system.…”

“…Note that a traditional wildfire detection framework using multi-modal data (Phan et al, 2020) would require an additional alignment step to align the timestamps of different data streams (e.g., via interpolation) so that each timestamp would have a full set of measurement values for wildfire detection. However, such an alignment step hinders the timeliness of detection since a window of a sufficient amount of data needs to be collected for an accurate alignment (Phan et al, 2020) (e.g., an interpolation often uses many neighbouring values to impute a missing value). Going beyond the state of the art, this alignment step is not needed in our framework thanks to our approach of complex event processing, which we will describe later.…”